tools/perf/builtin-stat.c

   1 /*
   2  * builtin-stat.c
   3  *
   4  * Builtin stat command: Give a precise performance counters summary
   5  * overview about any workload, CPU or specific PID.
   6  *
   7  * Sample output:
   8
   9    $ perf stat ~/hackbench 10
  10    Time: 0.104
  11
  12     Performance counter stats for '/home/mingo/hackbench':
  13
  14        1255.538611  task clock ticks     #      10.143 CPU utilization factor
  15              54011  context switches     #       0.043 M/sec
  16                385  CPU migrations       #       0.000 M/sec
  17              17755  pagefaults           #       0.014 M/sec
  18         3808323185  CPU cycles           #    3033.219 M/sec
  19         1575111190  instructions         #    1254.530 M/sec
  20           17367895  cache references     #      13.833 M/sec
  21            7674421  cache misses         #       6.112 M/sec
  22
  23     Wall-clock time elapsed:   123.786620 msecs
  24
  25  *
  26  * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
  27  *
  28  * Improvements and fixes by:
  29  *
  30  *   Arjan van de Ven <arjan@linux.intel.com>
  31  *   Yanmin Zhang <yanmin.zhang@intel.com>
  32  *   Wu Fengguang <fengguang.wu@intel.com>
  33  *   Mike Galbraith <efault@gmx.de>
  34  *   Paul Mackerras <paulus@samba.org>
  35  *
  36  * Released under the GPL v2. (and only v2, not any later version)
  37  */
  38
  39 #include "perf.h"
  40 #include "builtin.h"
  41 #include "util/util.h"
  42 #include "util/parse-options.h"
  43 #include "util/parse-events.h"
  44
  45 #include <sys/prctl.h>
  46 #include <math.h>
  47
  48 static struct perf_counter_attr default_attrs[MAX_COUNTERS] = {
  49
  50   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK      },
  51   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES},
  52   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS  },
  53   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS     },
  54
  55   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES      },
  56   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS    },
  57   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES},
  58   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES    },
  59
  60 };
  61
  62 static int                      system_wide                     =  0;
  63 static int                      inherit                         =  1;
  64 static int                      verbose                         =  0;
  65
  66 static int                      fd[MAX_NR_CPUS][MAX_COUNTERS];
  67
  68 static int                      target_pid                      = -1;
  69 static int                      nr_cpus                         =  0;
  70 static unsigned int             page_size;
  71
  72 static int                      scale                           =  1;
  73
  74 static const unsigned int default_count[] = {
  75         1000000,
  76         1000000,
  77           10000,
  78           10000,
  79         1000000,
  80           10000,
  81 };
  82
  83 #define MAX_RUN 100
  84
  85 static int                      run_count               =  1;
  86 static int                      run_idx                 =  0;
  87
  88 static u64                      event_res[MAX_RUN][MAX_COUNTERS][3];
  89 static u64                      event_scaled[MAX_RUN][MAX_COUNTERS];
  90
  91 //static u64                    event_hist[MAX_RUN][MAX_COUNTERS][3];
  92
  93
  94 static u64                      runtime_nsecs[MAX_RUN];
  95 static u64                      walltime_nsecs[MAX_RUN];
  96 static u64                      runtime_cycles[MAX_RUN];
  97
  98 static u64                      event_res_avg[MAX_COUNTERS][3];
  99 static u64                      event_res_noise[MAX_COUNTERS][3];
 100
 101 static u64                      event_scaled_avg[MAX_COUNTERS];
 102
 103 static u64                      runtime_nsecs_avg;
 104 static u64                      runtime_nsecs_noise;
 105
 106 static u64                      walltime_nsecs_avg;
 107 static u64                      walltime_nsecs_noise;
 108
 109 static u64                      runtime_cycles_avg;
 110 static u64                      runtime_cycles_noise;
 111
 112
 113 #define ERR_PERF_OPEN \
 114 "Error: counter %d, sys_perf_counter_open() syscall returned with %d (%s)\n"
 115
 116 static void create_perf_stat_counter(int counter)
 117 {
 118         struct perf_counter_attr *attr = attrs + counter;
 119
 120         if (scale)
 121                 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
 122                                     PERF_FORMAT_TOTAL_TIME_RUNNING;
 123
 124         if (system_wide) {
 125                 int cpu;
 126                 for (cpu = 0; cpu < nr_cpus; cpu++) {
 127                         fd[cpu][counter] = sys_perf_counter_open(attr, -1, cpu, -1, 0);
 128                         if (fd[cpu][counter] < 0 && verbose)
 129                                 fprintf(stderr, ERR_PERF_OPEN, counter,
 130                                         fd[cpu][counter], strerror(errno));
 131                 }
 132         } else {
 133                 attr->inherit   = inherit;
 134                 attr->disabled  = 1;
 135
 136                 fd[0][counter] = sys_perf_counter_open(attr, 0, -1, -1, 0);
 137                 if (fd[0][counter] < 0 && verbose)
 138                         fprintf(stderr, ERR_PERF_OPEN, counter,
 139                                 fd[0][counter], strerror(errno));
 140         }
 141 }
 142
 143 /*
 144  * Does the counter have nsecs as a unit?
 145  */
 146 static inline int nsec_counter(int counter)
 147 {
 148         if (attrs[counter].type != PERF_TYPE_SOFTWARE)
 149                 return 0;
 150
 151         if (attrs[counter].config == PERF_COUNT_SW_CPU_CLOCK)
 152                 return 1;
 153
 154         if (attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK)
 155                 return 1;
 156
 157         return 0;
 158 }
 159
 160 /*
 161  * Read out the results of a single counter:
 162  */
 163 static void read_counter(int counter)
 164 {
 165         u64 *count, single_count[3];
 166         ssize_t res;
 167         int cpu, nv;
 168         int scaled;
 169
 170         count = event_res[run_idx][counter];
 171
 172         count[0] = count[1] = count[2] = 0;
 173
 174         nv = scale ? 3 : 1;
 175         for (cpu = 0; cpu < nr_cpus; cpu++) {
 176                 if (fd[cpu][counter] < 0)
 177                         continue;
 178
 179                 res = read(fd[cpu][counter], single_count, nv * sizeof(u64));
 180                 assert(res == nv * sizeof(u64));
 181                 close(fd[cpu][counter]);
 182                 fd[cpu][counter] = -1;
 183
 184                 count[0] += single_count[0];
 185                 if (scale) {
 186                         count[1] += single_count[1];
 187                         count[2] += single_count[2];
 188                 }
 189         }
 190
 191         scaled = 0;
 192         if (scale) {
 193                 if (count[2] == 0) {
 194                         event_scaled[run_idx][counter] = -1;
 195                         count[0] = 0;
 196                         return;
 197                 }
 198
 199                 if (count[2] < count[1]) {
 200                         event_scaled[run_idx][counter] = 1;
 201                         count[0] = (unsigned long long)
 202                                 ((double)count[0] * count[1] / count[2] + 0.5);
 203                 }
 204         }
 205         /*
 206          * Save the full runtime - to allow normalization during printout:
 207          */
 208         if (attrs[counter].type == PERF_TYPE_SOFTWARE &&
 209                 attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK)
 210                 runtime_nsecs[run_idx] = count[0];
 211         if (attrs[counter].type == PERF_TYPE_HARDWARE &&
 212                 attrs[counter].config == PERF_COUNT_HW_CPU_CYCLES)
 213                 runtime_cycles[run_idx] = count[0];
 214 }
 215
 216 static int run_perf_stat(int argc, const char **argv)
 217 {
 218         unsigned long long t0, t1;
 219         int status = 0;
 220         int counter;
 221         int pid;
 222
 223         if (!system_wide)
 224                 nr_cpus = 1;
 225
 226         for (counter = 0; counter < nr_counters; counter++)
 227                 create_perf_stat_counter(counter);
 228
 229         /*
 230          * Enable counters and exec the command:
 231          */
 232         t0 = rdclock();
 233         prctl(PR_TASK_PERF_COUNTERS_ENABLE);
 234
 235         if ((pid = fork()) < 0)
 236                 perror("failed to fork");
 237
 238         if (!pid) {
 239                 if (execvp(argv[0], (char **)argv)) {
 240                         perror(argv[0]);
 241                         exit(-1);
 242                 }
 243         }
 244
 245         wait(&status);
 246
 247         prctl(PR_TASK_PERF_COUNTERS_DISABLE);
 248         t1 = rdclock();
 249
 250         walltime_nsecs[run_idx] = t1 - t0;
 251
 252         for (counter = 0; counter < nr_counters; counter++)
 253                 read_counter(counter);
 254
 255         return WEXITSTATUS(status);
 256 }
 257
 258 static void print_noise(u64 *count, u64 *noise)
 259 {
 260         if (run_count > 1)
 261                 fprintf(stderr, "   ( +- %7.3f%% )",
 262                         (double)noise[0]/(count[0]+1)*100.0);
 263 }
 264
 265 static void nsec_printout(int counter, u64 *count, u64 *noise)
 266 {
 267         double msecs = (double)count[0] / 1000000;
 268
 269         fprintf(stderr, " %14.6f  %-20s", msecs, event_name(counter));
 270
 271         if (attrs[counter].type == PERF_TYPE_SOFTWARE &&
 272                 attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK) {
 273
 274                 if (walltime_nsecs_avg)
 275                         fprintf(stderr, " # %10.3f CPUs ",
 276                                 (double)count[0] / (double)walltime_nsecs_avg);
 277         }
 278         print_noise(count, noise);
 279 }
 280
 281 static void abs_printout(int counter, u64 *count, u64 *noise)
 282 {
 283         fprintf(stderr, " %14Ld  %-20s", count[0], event_name(counter));
 284
 285         if (runtime_cycles_avg &&
 286                 attrs[counter].type == PERF_TYPE_HARDWARE &&
 287                         attrs[counter].config == PERF_COUNT_HW_INSTRUCTIONS) {
 288
 289                 fprintf(stderr, " # %10.3f IPC  ",
 290                         (double)count[0] / (double)runtime_cycles_avg);
 291         } else {
 292                 if (runtime_nsecs_avg) {
 293                         fprintf(stderr, " # %10.3f M/sec",
 294                                 (double)count[0]/runtime_nsecs_avg*1000.0);
 295                 }
 296         }
 297         print_noise(count, noise);
 298 }
 299
 300 /*
 301  * Print out the results of a single counter:
 302  */
 303 static void print_counter(int counter)
 304 {
 305         u64 *count, *noise;
 306         int scaled;
 307
 308         count = event_res_avg[counter];
 309         noise = event_res_noise[counter];
 310         scaled = event_scaled_avg[counter];
 311
 312         if (scaled == -1) {
 313                 fprintf(stderr, " %14s  %-20s\n",
 314                         "<not counted>", event_name(counter));
 315                 return;
 316         }
 317
 318         if (nsec_counter(counter))
 319                 nsec_printout(counter, count, noise);
 320         else
 321                 abs_printout(counter, count, noise);
 322
 323         if (scaled)
 324                 fprintf(stderr, "  (scaled from %.2f%%)",
 325                         (double) count[2] / count[1] * 100);
 326
 327         fprintf(stderr, "\n");
 328 }
 329
 330 /*
 331  * normalize_noise noise values down to stddev:
 332  */
 333 static void normalize_noise(u64 *val)
 334 {
 335         double res;
 336
 337         res = (double)*val / (run_count * sqrt((double)run_count));
 338
 339         *val = (u64)res;
 340 }
 341
 342 static void update_avg(const char *name, int idx, u64 *avg, u64 *val)
 343 {
 344         *avg += *val;
 345
 346         if (verbose > 1)
 347                 fprintf(stderr, "debug: %20s[%d]: %Ld\n", name, idx, *val);
 348 }
 349 /*
 350  * Calculate the averages and noises:
 351  */
 352 static void calc_avg(void)
 353 {
 354         int i, j;
 355
 356         if (verbose > 1)
 357                 fprintf(stderr, "\n");
 358
 359         for (i = 0; i < run_count; i++) {
 360                 update_avg("runtime", 0, &runtime_nsecs_avg, runtime_nsecs + i);
 361                 update_avg("walltime", 0, &walltime_nsecs_avg, walltime_nsecs + i);
 362                 update_avg("runtime_cycles", 0, &runtime_cycles_avg, runtime_cycles + i);
 363
 364                 for (j = 0; j < nr_counters; j++) {
 365                         update_avg("counter/0", j,
 366                                 event_res_avg[j]+0, event_res[i][j]+0);
 367                         update_avg("counter/1", j,
 368                                 event_res_avg[j]+1, event_res[i][j]+1);
 369                         update_avg("counter/2", j,
 370                                 event_res_avg[j]+2, event_res[i][j]+2);
 371                         update_avg("scaled", j,
 372                                 event_scaled_avg + j, event_scaled[i]+j);
 373                 }
 374         }
 375         runtime_nsecs_avg /= run_count;
 376         walltime_nsecs_avg /= run_count;
 377         runtime_cycles_avg /= run_count;
 378
 379         for (j = 0; j < nr_counters; j++) {
 380                 event_res_avg[j][0] /= run_count;
 381                 event_res_avg[j][1] /= run_count;
 382                 event_res_avg[j][2] /= run_count;
 383         }
 384
 385         for (i = 0; i < run_count; i++) {
 386                 runtime_nsecs_noise +=
 387                         abs((s64)(runtime_nsecs[i] - runtime_nsecs_avg));
 388                 walltime_nsecs_noise +=
 389                         abs((s64)(walltime_nsecs[i] - walltime_nsecs_avg));
 390                 runtime_cycles_noise +=
 391                         abs((s64)(runtime_cycles[i] - runtime_cycles_avg));
 392
 393                 for (j = 0; j < nr_counters; j++) {
 394                         event_res_noise[j][0] +=
 395                                 abs((s64)(event_res[i][j][0] - event_res_avg[j][0]));
 396                         event_res_noise[j][1] +=
 397                                 abs((s64)(event_res[i][j][1] - event_res_avg[j][1]));
 398                         event_res_noise[j][2] +=
 399                                 abs((s64)(event_res[i][j][2] - event_res_avg[j][2]));
 400                 }
 401         }
 402
 403         normalize_noise(&runtime_nsecs_noise);
 404         normalize_noise(&walltime_nsecs_noise);
 405         normalize_noise(&runtime_cycles_noise);
 406
 407         for (j = 0; j < nr_counters; j++) {
 408                 normalize_noise(&event_res_noise[j][0]);
 409                 normalize_noise(&event_res_noise[j][1]);
 410                 normalize_noise(&event_res_noise[j][2]);
 411         }
 412 }
 413
 414 static void print_stat(int argc, const char **argv)
 415 {
 416         int i, counter;
 417
 418         calc_avg();
 419
 420         fflush(stdout);
 421
 422         fprintf(stderr, "\n");
 423         fprintf(stderr, " Performance counter stats for \'%s", argv[0]);
 424
 425         for (i = 1; i < argc; i++)
 426                 fprintf(stderr, " %s", argv[i]);
 427
 428         fprintf(stderr, "\'");
 429         if (run_count > 1)
 430                 fprintf(stderr, " (%d runs)", run_count);
 431         fprintf(stderr, ":\n\n");
 432
 433         for (counter = 0; counter < nr_counters; counter++)
 434                 print_counter(counter);
 435
 436
 437         fprintf(stderr, "\n");
 438         fprintf(stderr, " %14.9f  seconds time elapsed.\n",
 439                         (double)walltime_nsecs_avg/1e9);
 440         fprintf(stderr, "\n");
 441 }
 442
 443 static volatile int signr = -1;
 444
 445 static void skip_signal(int signo)
 446 {
 447         signr = signo;
 448 }
 449
 450 static void sig_atexit(void)
 451 {
 452         if (signr == -1)
 453                 return;
 454
 455         signal(signr, SIG_DFL);
 456         kill(getpid(), signr);
 457 }
 458
 459 static const char * const stat_usage[] = {
 460         "perf stat [<options>] <command>",
 461         NULL
 462 };
 463
 464 static const struct option options[] = {
 465         OPT_CALLBACK('e', "event", NULL, "event",
 466                      "event selector. use 'perf list' to list available events",
 467                      parse_events),
 468         OPT_BOOLEAN('i', "inherit", &inherit,
 469                     "child tasks inherit counters"),
 470         OPT_INTEGER('p', "pid", &target_pid,
 471                     "stat events on existing pid"),
 472         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 473                             "system-wide collection from all CPUs"),
 474         OPT_BOOLEAN('S', "scale", &scale,
 475                             "scale/normalize counters"),
 476         OPT_BOOLEAN('v', "verbose", &verbose,
 477                     "be more verbose (show counter open errors, etc)"),
 478         OPT_INTEGER('r', "repeat", &run_count,
 479                     "repeat command and print average + stddev (max: 100)"),
 480         OPT_END()
 481 };
 482
 483 int cmd_stat(int argc, const char **argv, const char *prefix)
 484 {
 485         int status;
 486
 487         page_size = sysconf(_SC_PAGE_SIZE);
 488
 489         memcpy(attrs, default_attrs, sizeof(attrs));
 490
 491         argc = parse_options(argc, argv, options, stat_usage, 0);
 492         if (!argc)
 493                 usage_with_options(stat_usage, options);
 494         if (run_count <= 0 || run_count > MAX_RUN)
 495                 usage_with_options(stat_usage, options);
 496
 497         if (!nr_counters)
 498                 nr_counters = 8;
 499
 500         nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
 501         assert(nr_cpus <= MAX_NR_CPUS);
 502         assert(nr_cpus >= 0);
 503
 504         /*
 505          * We dont want to block the signals - that would cause
 506          * child tasks to inherit that and Ctrl-C would not work.
 507          * What we want is for Ctrl-C to work in the exec()-ed
 508          * task, but being ignored by perf stat itself:
 509          */
 510         atexit(sig_atexit);
 511         signal(SIGINT,  skip_signal);
 512         signal(SIGALRM, skip_signal);
 513         signal(SIGABRT, skip_signal);
 514
 515         status = 0;
 516         for (run_idx = 0; run_idx < run_count; run_idx++) {
 517                 if (run_count != 1 && verbose)
 518                         fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx+1);
 519                 status = run_perf_stat(argc, argv);
 520         }
 521
 522         print_stat(argc, argv);
 523
 524         return status;
 525 }